01 — 60-SECOND READ
The Skim
01
OpenAI plans ChatGPT overhaul: 'Chat is dead,' pivot to agent superapp with coding, tools, integrations. OpenAI is rebuilding ChatGPT as an agent OS, not a chat interface—a fundamental shift toward autonomous task execution that mirrors 16x9's thesis on AI-native operating models and agent-first architecture.
→ us: This is the market admitting what 16x9 ships: chat is a crutch. The future is agents that run tasks without you. ChatGPT becoming a superapp validates the move from 'talk to AI' to 'AI runs your company.'
02
Anthropic hires OpenAI's second chip engineer as both race IPO and custom silicon. Both companies are now building custom silicon in-house; Anthropic's chip push dramatically lowers inference costs and reduces dependence on NVIDIA, a critical move ahead of IPO.
→ us: 16x9 runs on Claude and depends on Anthropic's efficiency roadmap: custom chips are the play to slash token costs, extend margin, and ship agent infrastructure cheaper than competitors.
03
Perplexity launches 'Search as Code': agents write own search pipelines, cut token costs 85%, beat OpenAI and Anthropic on benchmarks. Agent can now write and execute its own search logic in Python instead of calling fixed APIs; massive token efficiency gain and benchmark win reshapes how retrieval-augmented reasoning will be built.
→ us: This is the move from API integration to agent autonomy: agents stop calling tools, they write tools. Cost per token drops 85% because the model isn't chatting—it's coding its own retrieval. 16x9's thesis on durable, self-improving workflows.
04
Huawei-led team completes post-training of DeepSeek V4-Pro 1.6T model on 1,000 Ascend chips. China has now demonstrated end-to-end large-scale model training without NVIDIA; proves Ascend silicon is viable for trillion-parameter scale, reshaping the geopolitical chip economy.
→ us: The US chip export regime is cracking: China has shipped a production-grade trillion-parameter model trained entirely on domestic silicon, compressing the compute advantage that underpins American AI leadership.
05
AI power demand accelerating faster than energy supply; grid strain intensifies. Data centers are consuming more power than grid capacity can supply; energy is now the hard constraint on AI scaling, not chips or models.
→ us: 16x9 scales on Claude, which scales on power. The constraint is no longer NVIDIA—it's kilowatts. Every efficiency gain (smaller models, better inference, agent orchestration) compounds into leverage on the most scarce resource: grid capacity.
06
OpenAI ships ChatGPT Lockdown Mode: disables web, Deep Research, Agent Mode to block prompt injection exfiltration. Prompt injection remains an unsolved problem; Lockdown Mode is a partial mitigation (blocks final exfiltration step) but does not prevent the core attack, raising questions about agent safety at scale.
07
Sriram Krishnan leaves White House AI advisor role to launch new policy institution. Key Trump administration AI policy architect is exiting to build a new think tank; signals shift in how AI governance will be shaped outside formal government.
02 — WHERE THE ACTION SITS
The 5 Layers
Energy01
Data centers are consuming more power than grid capacity can supply; energy is now the hard constraint on AI scaling, not chips or models.
News
→ us: 16x9 scales on Claude, which scales on power. The constraint is no longer NVIDIA—it's kilowatts. Every efficiency gain (smaller models, better inference, agent orchestration) compounds into leverage on the most scarce resource: grid capacity.
Google News
→
Chips02
Both companies are now building custom silicon in-house; Anthropic's chip push dramatically lowers inference costs and reduces dependence on NVIDIA, a critical move ahead of IPO.
Clive Chan was OpenAI's second hardware employee; both companies preparing IPOs.News
→ us: 16x9 runs on Claude and depends on Anthropic's efficiency roadmap: custom chips are the play to slash token costs, extend margin, and ship agent infrastructure cheaper than competitors.
The Decoder · Google News
China has now demonstrated end-to-end large-scale model training without NVIDIA; proves Ascend silicon is viable for trillion-parameter scale, reshaping the geopolitical chip economy.
1.6 trillion parameters, 1,000 Ascend 910C chips, full post-training completed.Research
→ us: The US chip export regime is cracking: China has shipped a production-grade trillion-parameter model trained entirely on domestic silicon, compressing the compute advantage that underpins American AI leadership.
Tom's Hardware · Google News
→
Data Centers03
— quiet today
→
Models04
— quiet today
→
Applications05
OpenAI is rebuilding ChatGPT as an agent OS, not a chat interface—a fundamental shift toward autonomous task execution that mirrors 16x9's thesis on AI-native operating models and agent-first architecture.
'Chat is dead' — internal OpenAI statement.News
→ us: This is the market admitting what 16x9 ships: chat is a crutch. The future is agents that run tasks without you. ChatGPT becoming a superapp validates the move from 'talk to AI' to 'AI runs your company.'
The Decoder
Agent can now write and execute its own search logic in Python instead of calling fixed APIs; massive token efficiency gain and benchmark win reshapes how retrieval-augmented reasoning will be built.
85% token cost reduction; outperforms OpenAI and Anthropic on benchmarks.Launch
→ us: This is the move from API integration to agent autonomy: agents stop calling tools, they write tools. Cost per token drops 85% because the model isn't chatting—it's coding its own retrieval. 16x9's thesis on durable, self-improving workflows.
The Decoder
Prompt injection remains an unsolved problem; Lockdown Mode is a partial mitigation (blocks final exfiltration step) but does not prevent the core attack, raising questions about agent safety at scale.
Disables web access, Deep Research, Agent Mode; does not fully prevent prompt injection attacks.Launch
The Decoder
03 — TO USE · TO WATCH
Tools & Horizon
Tools you can use today
- Perplexity launches 'Search as Code': agents write own search pipelines, cut token costs 85%, beat OpenAI and Anthropic on benchmarks Applications — Agent can now write and execute its own search logic in Python instead of calling fixed APIs; massive token efficiency gain and benchmark win reshapes how retrieval-augmented reasoning will be built.
- OpenAI ships ChatGPT Lockdown Mode: disables web, Deep Research, Agent Mode to block prompt injection exfiltration Applications — Prompt injection remains an unsolved problem; Lockdown Mode is a partial mitigation (blocks final exfiltration step) but does not prevent the core attack, raising questions about agent safety at scale.
On the horizon
- nothing today
04 — THE WIRE · LATEST FIRST
Headlines
11:37zAssessing Solventum (SOLV) Valuation As AI Disruption Fears Challenge Health Information Systems Prospects - simplywall.stAI markets — IPO/funding/earnings (Google News)
11:33zDeveloper gets Half-Life running at 30 FPS on a Nokia N95 — proves 2007 phones can just about match 1998 PCsTom's Hardware
11:32zThe data center backlash coming to key midterm battlegrounds - The Washington PostAI energy / power / nuclear (Google News)
11:31zSpaceX vs Anthropic: Here's Which Mega-IPO I'm More Likely to Buy - The Globe and MailAnthropic (Google News)
11:20zAI companies are barreling toward huge Wall Street debuts. A look at the biggest players, Digital News - AsiaOneAI markets — IPO/funding/earnings (Google News)
11:09zAIQ turned $10,000 into $13,400 in six months as AI chips soared 34% YTD - MSNAI markets — IPO/funding/earnings (Google News)
11:00zLudicrous overclock slams 1.7 volts into 6700K in an attempt to stop CPU from bottlenecking an RTX 3080 — 5.2 GHz on aging four-core pushes GPU utilization from 60% to 74%Tom's Hardware
10:57zMore than 100 campaigners back MP suing Elon Musk’s xAI over deepfake images of her - The IndependentxAI / Grok (Google News)
10:50zAnthropic’s Claude Opus 4.7 matches dedicated NMR software in chemistry tasks - Crypto BriefingAnthropic (Google News)
10:31zExclusive / Saudi’s NEOM faces $16 billion bill to cancel NEOM contracts - SemaforSemafor — AI/tech (Google News)